Spaces:
Runtime error
Runtime error
File size: 6,605 Bytes
a5f8a35 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 |
import bisect
import math
from typing import List
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LambdaLR
class LinearWarmupNoDecayLR(LambdaLR):
r"""
A learning rate scheduler which linearly increases learning rate from 0
LR, and further keeps it constant throughout training.
Parameters
----------
optimizer: torch.optim.Optimizer
Wrapper optimizer.
total_steps: int
Total epochs (or iterations) for training.
warmup_steps: int
Number of first few steps to do linear warmup.
last_epoch: int, optional (default = -1)
The index of last step (epoch or iteration). We named it ``last_epoch``
instead of ``last_step`` to keep the naming consistent with other LR
schedulers in PyTorch.
"""
def __init__(
self,
optimizer: Optimizer,
total_steps: int,
warmup_steps: int,
last_epoch: int = -1,
):
assert (
warmup_steps < total_steps
), "Warmup steps should be less than total steps."
self.tsteps = total_steps
self.wsteps = warmup_steps
super().__init__(optimizer, self._lr_multiplier, last_epoch)
def _lr_multiplier(self, step: int) -> float:
multiplier = step / float(max(1, self.wsteps)) if step < self.wsteps else 1
return max(0, multiplier)
class LinearWarmupMultiStepLR(LambdaLR):
r"""
A learning rate scheduler which linearly increases learning rate from 0
LR, and further decreases it by gamma once the number of steps reaches one
of the milestones.
Parameters
----------
optimizer: torch.optim.Optimizer
Wrapper optimizer.
total_steps: int
Total epochs (or iterations) for training.
warmup_steps: int
Number of first few steps to do linear warmup.
milestones: List[int]
List of step indices (epochs or iterations depending on context). Must
be increasing.
gamma: float, optional (default = 0.1)
Multiplicative factor of learning rate decay.
last_epoch: int, optional (default = -1)
The index of last step (epoch or iteration). We named it ``last_epoch``
instead of ``last_step`` to keep the naming consistent with other LR
schedulers in PyTorch.
"""
def __init__(
self,
optimizer: Optimizer,
total_steps: int,
warmup_steps: int,
milestones: List[int],
gamma: float = 0.1,
last_epoch: int = -1,
):
self.wsteps = warmup_steps
self.milestones = milestones
self.gamma = gamma
# Keep a track of number of milestones encountered.
self.milestones_so_far = 0
# Common sanity checks.
assert milestones == sorted(milestones), "milestones must be increasing"
assert milestones[0] > warmup_steps, "first milestone must be after warmup"
assert (
milestones[-1] < total_steps
), "last milestone must be less than total steps"
super().__init__(optimizer, self._lr_multiplier, last_epoch)
def _lr_multiplier(self, step: int) -> float:
if step < self.wsteps:
# Linear warmup.
multiplier = step / float(max(1, self.wsteps))
else:
# Step decay based on milestones.
multiplier = self.gamma ** bisect.bisect_right(self.milestones, step)
# Avoid negative learning rate.
return max(0, multiplier)
class LinearWarmupLinearDecayLR(LambdaLR):
r"""
A learning rate scheduler which linearly increases learning rate from 0
LR, and further decreases it linearly to zero.
Parameters
----------
optimizer: torch.optim.Optimizer
Wrapper optimizer.
total_steps: int
Total epochs (or iterations) for training.
warmup_steps: int
Number of first few steps to do linear warmup.
last_epoch: int, optional (default = -1)
The index of last step (epoch or iteration). We named it ``last_epoch``
instead of ``last_step`` to keep the naming consistent with other LR
schedulers in PyTorch.
"""
def __init__(
self,
optimizer: Optimizer,
total_steps: int,
warmup_steps: int,
last_epoch: int = -1,
):
assert (
warmup_steps < total_steps
), "Warmup steps should be less than total steps."
self.tsteps = total_steps
self.wsteps = warmup_steps
super().__init__(optimizer, self._lr_multiplier, last_epoch)
def _lr_multiplier(self, step: int) -> float:
if step < self.wsteps:
# Linear warmup.
multiplier = step / float(max(1, self.wsteps))
else:
# Linear decay.
multiplier = (self.tsteps - step) / (self.tsteps - self.wsteps)
# Avoid negative learning rate.
return max(0, multiplier)
class LinearWarmupCosineAnnealingLR(LambdaLR):
r"""
A learning rate scheduler which linearly increases learning rate from 0
LR, and further decreases it to zero by cosine decay. After linear warmup,
the LR decays as:
.. math::
\eta_t = \eta_{max}\cos^2(\frac{T_{cur} - T_{warm}}{T_{max} - T_{warm}}\frac{\pi}{2})
Parameters
----------
optimizer: torch.optim.Optimizer
Wrapper optimizer.
total_steps: int
Total epochs (or iterations) for training.
warmup_steps: int
Number of first few steps to do linear warmup.
last_epoch: int, optional (default = -1)
The index of last step (epoch or iteration). We named it ``last_epoch``
instead of ``last_step`` to keep the naming consistent with other LR
schedulers in PyTorch.
"""
def __init__(
self,
optimizer: Optimizer,
total_steps: int,
warmup_steps: int,
last_epoch: int = -1,
):
assert (
warmup_steps < total_steps
), "Warmup steps should be less than total steps."
self.tsteps = total_steps
self.wsteps = warmup_steps
super().__init__(optimizer, self._lr_multiplier, last_epoch)
def _lr_multiplier(self, step: int) -> float:
if step < self.wsteps:
# Linear warmup.
multiplier = step / float(max(1, self.wsteps))
else:
# Cosine annealing decay.
cos_factor = (step - self.wsteps) / (self.tsteps - self.wsteps)
multiplier = math.cos(cos_factor * (math.pi / 2)) ** 2
# Avoid negative learning rate.
return max(0, multiplier)
|